On Sun, Aug 18, 2019 at 08:04:11AM +0800, Gao Xiang wrote:
> On Sun, Aug 18, 2019 at 07:38:47AM +0800, Gao Xiang wrote:
> > Hi Richard,
> >
> > On Sun, Aug 18, 2019 at 01:25:58AM +0200, Richard Weinberger wrote:
>
> []
>
> > >
> > > While digging a little into the code I noticed that you have very few
> > > checks of the on-disk data.
> > > For example ->u.i_blkaddr. I gave it a try and created a
> > > malformed filesystem where u.i_blkaddr is 0xdeadbeef, it causes the kernel
> > > to loop forever around erofs_read_raw_page().
> >
> > I don't fuzz all the on-disk fields for EROFS, I will do later..
> > You can see many in-kernel filesystems are still hardening the related
> > stuff. Anyway, I will dig into this field you mentioned recently, but
> > I think it can be fixed easily later.
>
> ...I take a simple try with the following erofs-utils diff and
> a directory containing enwik9 only, with the latest kernel (5.3-rc)
> and command line is
> mkfs/mkfs.erofs -d9 enwik9.img testdir.
>
> diff --git a/lib/inode.c b/lib/inode.c
> index 581f263..2540338 100644
> --- a/lib/inode.c
> +++ b/lib/inode.c
> @@ -388,8 +388,7 @@ static bool erofs_bh_flush_write_inode(struct erofs_buffer_head *bh)
> v1.i_u.compressed_blocks =
> cpu_to_le32(inode->u.i_blocks);
> else
> - v1.i_u.raw_blkaddr =
> - cpu_to_le32(inode->u.i_blkaddr);
> + v1.i_u.raw_blkaddr = 0xdeadbeef;
> break;
> }
>
> I tested the corrupted image with looped device and real blockdevice
> by dd, and it seems fine....
> [36283.012381] erofs: initializing erofs 1.0
> [36283.012510] erofs: successfully to initialize erofs
> [36283.012975] erofs: read_super, device -> /dev/loop17
> [36283.012976] erofs: options -> (null)
> [36283.012983] erofs: root inode @ nid 36
> [36283.012995] erofs: mounted on /dev/loop17 with opts: (null).
> [36297.354090] attempt to access beyond end of device
> [36297.354098] loop17: rw=0, want=29887428984, limit=1953128
> [36297.354107] attempt to access beyond end of device
> [36297.354109] loop17: rw=0, want=29887428480, limit=1953128
> [36301.827234] attempt to access beyond end of device
> [36301.827243] loop17: rw=0, want=29887428480, limit=1953128
> [36371.426889] erofs: unmounted for /dev/loop17
> [36518.156114] erofs: read_super, device -> /dev/nvme0n1p4
> [36518.156115] erofs: options -> (null)
> [36518.156260] erofs: root inode @ nid 36
> [36518.156384] erofs: mounted on /dev/nvme0n1p4 with opts: (null).
> [36522.818884] attempt to access beyond end of device
> [36522.818889] nvme0n1p4: rw=0, want=29887428984, limit=62781440
> [36522.818895] attempt to access beyond end of device
> [36522.818896] nvme0n1p4: rw=0, want=29887428480, limit=62781440
> [36524.072018] attempt to access beyond end of device
> [36524.072028] nvme0n1p4: rw=0, want=29887428480, limit=62781440
>
> Could you give me more hints how to reproduce that? and I will
> dig into more maybe it needs more conditions...
I think I found what happened here... That is not a bug due to lack of
check of on-disk ->u.i_blkaddr (seems block layer will handle access
beyond end of device) but actually a bug of erofs_readdir:
diff --git a/fs/erofs/data.c b/fs/erofs/data.c
index fda16ec8863e..5b5f35d47370 100644
--- a/fs/erofs/data.c
+++ b/fs/erofs/data.c
@@ -329,6 +329,8 @@ static int erofs_raw_access_readpage(struct file *file, struct page *page)
trace_erofs_readpage(page, true);
+ WARN_ON(1);
+
bio = erofs_read_raw_page(NULL, page->mapping,
page, &last_block, 1, false);
@@ -379,6 +381,8 @@ static int erofs_raw_access_readpages(struct file *filp,
/* the rare case (end in gaps) */
if (unlikely(bio))
__submit_bio(bio, REQ_OP_READ, 0);
+
+ WARN_ON(1);
return 0;
}
diff --git a/fs/erofs/dir.c b/fs/erofs/dir.c
index 637d70108d59..ccca954438ed 100644
--- a/fs/erofs/dir.c
+++ b/fs/erofs/dir.c
@@ -80,8 +80,10 @@ static int erofs_readdir(struct file *f, struct dir_context *ctx)
unsigned int nameoff, maxsize;
dentry_page = read_mapping_page(mapping, i, NULL);
- if (IS_ERR(dentry_page))
- continue;
+ if (IS_ERR(dentry_page)) {
+ err = PTR_ERR(dentry_page);
+ break;
+ }
de = (struct erofs_dirent *)kmap(dentry_page);
It's a forever loop due to error handling of the read_mapping_page above.
I will fix that in another patch and thanks for your report!
Thanks,
Gao Xiang
>
> Thanks,
> Gao Xiang
>
> >
> > Thanks,
> > Gao Xiang
> >
> > >
> > > Thanks,
> > > //richard
